SOV-Transcribe-1 is an open-weight speech recognition model released by Salesforce on November 12, 2026, topping the Artificial Analysis open-source leaderboard two days after launch. Here is the verdict: its accuracy is genuinely the best in open-source transcription, but the 53GB weights and the 96GB VRAM test setup shut out almost all individual users. If you do not have a large-VRAM GPU, just watch from the sidelines for now — there is no rush to try it.
What the 2.28% Top Spot Is Really Worth
The architecture pairs a Whisper large-v3 encoder and projection layer with a Gemma 4 26B A4B MoE decoder, at 26.6B total parameters, doing non-streaming recognition. On the AA-WER leaderboard its word error rate is 2.28%, first among open-weight models and seventh out of all 44 models, ahead of Nemotron 3.5 Transcribe at 2.81% and Whisper Large v3 Turbo at 4.62%. It lets you control transcription style in natural language, and it is licensed under Apache 2.0, so commercial use is allowed. The limits are clear too: it supports only seven European languages, no Chinese, and a single audio clip of at most 30 minutes.
Deployment Cost: 96GB of VRAM Is the Hard Floor
The official repository lives on Hugging Face, under the organization salesforce-sovereign-intelligence, project SOV-Transcribe-1; the weights and code are all on that model page, and there is no separate GitHub repository. There are two ways to deploy it: the bundled vLLM 0.32.0 plugin, which can serve an OpenAI-compatible transcription service, or Transformers 5.19 to 5.22 with the repository's own modeling code, which requires trust_remote_code. The cost is the real issue: the weights are about 53GB, so 12GB, 16GB and 24GB consumer cards cannot hold them, and the official numbers were measured on a single 96GB RTX PRO 6000. vLLM's default of 1,024 concurrent sequences will run out of memory, so max_num_seqs has to be lowered to 64. Speed is decent: a 30-second recording takes about 0.22 seconds to transcribe.
Pitfalls and Limits: Who Should Bother, and Who Should Not
First, the pitfalls: it is a non-streaming model, suited only to batch transcription of recorded meetings, calls and podcasts — not live captions or livestreams. As of now there is no official quantized or GGUF version, so getting it onto a consumer card means waiting for community quantization. A vLLM endpoint defaults to a maximum of 600 seconds of audio per request and needs reconfiguration to reach the 30-minute ceiling, and the model card itself notes occasional repeated output that needs post-processing to remove. Some benchmark data is not public, so the leaderboard scores cannot be fully reproduced externally. It fits organizations with 80GB-class GPUs and audio that must stay inside their own network, or teams whose transcription volume is large enough that self-hosting beats paying per API call. Individual users, Chinese transcription needs and live-caption scenarios are better off waiting — Whisper Large v3 Turbo is the easier option.
